Auto Decision Engine

Alpha prototype · in development

Decide whether a used car is worth pursuing, and what to check first.

For cautious buyers in Poland, Auto Decision Engine ties evidence to one specific listing, keeps the seller’s claims apart from what the listing shows, and turns what is still unknown into a prioritized next step.

Inside the engine Contact the founder

Illustrative example · fictional

Fictional listing: used petrol estate

Buyer contextNeeds reliable daily use for several years.

Listing text Reported
Seller says the engine was replaced after water entered the intake.
Not in the listing Unknown
No repair documents. Repair scope and current condition are unknown.

Not inferred: a flooded cabin, a current fault or a successful repair.

First next step

Request dated repair invoices and the scope of work, showing this car’s vehicle identifier.

WhyThis reported engine event makes its repair records more useful than a generic history check. Records cannot establish current condition.

If documents match this car
Compare their date and scope with the seller’s account, then arrange an independent inspection, including the engine and intake.
If missing or conflicting
Clarify before any deposit; the repair scope stays unknown.
Fictional illustration of the intended reasoning; not a real listing or generated output.

Inside the engine

Evidence first, then reasoning, then a checked report

The prototype combines multimodal AI, which reads listing text and photos, with explicit rules and source checks that decide what the evidence can support.

  1. Link observations to their sources

    Text and photos become observations, each linked to its source. Identifiers within the listing are checked for conflicts. Documents visible in photos stay unverified and are not treated as linked to the car. Equipment shown is not proof that it works.

  2. Keep claims, observations and knowledge apart

    Seller claims stay separate from what the listing shows. Model knowledge is checked against its version, engine and market restrictions; unresolved matches remain conditional. Comparable asking prices are used only when enough suitable offers exist.

  3. Prioritize conditional next steps

    The evidence sets which check comes first and why, and what the buyer should do if the answer goes either way, as in the example above. Missing evidence stays unknown rather than being filled in.

  4. Check the report against the evidence

    A drafted report is checked against its evidence. Flagged statements are corrected or withheld, and unresolved limitations stay visible. This checks support; it does not make the output true or complete.

These mechanisms run in the local prototype, and its deterministic parts are tested. Accuracy on new cases has not yet been measured. The engine does not replace a viewing, a history check or a mechanic.

Where it stands

An early alpha, built carefully

The project started in 2026 and runs as a local prototype. The goal is a service that takes a public listing link and returns an evidence-bound decision report for that exact car. So far it has been tried on retained cases and manually assisted reports. It is not yet an autonomous or public service, and this page offers no analysis.

Contact

Talk to the founder

If you buy, inspect or advise on used cars in Poland, or want to discuss the project, I would be glad to hear from you.

Karol, project founder